TrafficInternVL: Spatially-Guided Fine-Tuning with Caption Refinement for Fine-Grained Traffic Safety Captioning and Visual Question Answering
1
Issued Date
2025-01-01
Resource Type
Scopus ID
2-s2.0-105035187032
Journal Title
Proceedings 2025 IEEE Cvf International Conference on Computer Vision Workshops Iccv W 2025
Start Page
5358
End Page
5365
Rights Holder(s)
SCOPUS
Bibliographic Citation
Proceedings 2025 IEEE Cvf International Conference on Computer Vision Workshops Iccv W 2025 (2025) , 5358-5365
Suggested Citation
Phimsiri S., Sunpawatr S., Cherdchusakulchai R., Kiawjak P., Tosawadi T., Tungjitnob S., Trairattanapa V., Vatathanavaro S., Kudisthalert W., Utintu C., Saetan W., Kongsawat N., Borisuitsawat P., Mahakijdechachai K., Su-Inn N., Thamwiwatthana E., Suttichaya V. TrafficInternVL: Spatially-Guided Fine-Tuning with Caption Refinement for Fine-Grained Traffic Safety Captioning and Visual Question Answering. Proceedings 2025 IEEE Cvf International Conference on Computer Vision Workshops Iccv W 2025 (2025) , 5358-5365. 5365. doi:10.1109/ICCVW69036.2025.00559 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/116233
Title
TrafficInternVL: Spatially-Guided Fine-Tuning with Caption Refinement for Fine-Grained Traffic Safety Captioning and Visual Question Answering
Corresponding Author(s)
Other Contributor(s)
Abstract
Fine-grained traffic understanding requires both detailed visual descriptions and precise answers to safety-critical questions. We present TrafficInternVl, a framework for fine-grained traffic safety description and question answering, developed for AI City Challenge 2025 Track 2. Our approach is based on the InternVL3-38B vision-language model and integrates four key components: (1) spatially guided visual prompting via bounding-box-based cropping and rendering; (2) Adaptive view selection protocols; (3) low-rank adaptation (LoRA) fine-tuning, updating only 1% of model parameters; and (4) caption refinement for intra-scene consistency. Our model achieves a Caption Score of 32.75 (BLEU-4, METEOR, ROUGE-L, CIDEr averaged) and a VQA accuracy of 83.08 %. Code, prompts, and LoRA weights are released at https://github.com/ARV-MLCORE/TrafficInternVL